An AI assistant linked to the company's email and files goes through a "connector" that, as a rule, no one checks.
Connectors (so-called MCP servers) are programs that allow an AI assistant to access data from Gmail, Drive, M365, or a company’s database. The security firm OX Security analyzed 15,465 such public connectors from five online stores and reached a simple conclusion: the stores don’t verify anyone. Anyone can publish a connector without any oversight.
The figures show why this matters: 15.6% of connectors run on servers outside the U.S., some in China and Russia; 2.3% use expired domains, which anyone can buy for a few dollars and take over the connector. Worse still, a “remote” connector may run code that’s completely different from what you see publicly, so reviewing the code doesn’t tell you what it’s actually doing.
The risk for a small business is clear: a poorly written or compromised connector can see everything the assistant sees—all the emails and all the files the assistant has access to. Nothing needs to be “hacked”; you’ve handed over the key yourself.
What you need to do:
- Make a list of the connectors linked to the company’s AI tools and note which data sources they access: email, documents, accounting.
- Determine who in the company is allowed to link a work account to an AI tool—not everyone.
- Use only connectors from the service’s official provider, not from third-party repositories.
- Treat AI connectors like any other vendor with access to your data: keep them to a minimum, verify them, and grant them only the minimum necessary permissions.
Sources: The Hacker News (OX Security report) | https://thehackernews.com/2026/10/welcome-to-jungle-what-we-found-inside.html
This article was generated with AI assistance.